How do AI assistants choose their sources?

Modern AI assistants answer product questions in three steps: they issue live web searches, read a handful of returned pages, and synthesize one answer with citations attached. Which sources survive that filter is decided per question, but in practice engines favour pages that answer the question directly and come from a source their retrieval already trusts — comparison articles, review directories, community threads, documentation, and authoritative brand pages.

Last updated 2026-07-28

What happens between the question and the answer?

The assistant rewrites your question into one or more search queries, retrieves results, reads a small number of them, and generates an answer constrained to what it read — attaching citations to the pages it used. The number of pages actually read is small, usually a handful, which is why this is a winner-take-most surface rather than a ranked list.

The consequential part is that retrieval, not generation, does most of the filtering. If your page is not retrieved for the rewritten query, no amount of persuasive copy on it matters. This is why AEO work concentrates on being present in the sources that get retrieved for your category, and why the grounding mechanism is the thing to understand first.

What kinds of pages actually get cited?

For discovery questions, third-party pages dominate; for brand-specific questions, your own pages do. The pattern is consistent enough across engines to plan around.

• Comparison and "best of" articles — the single most common citation type for discovery questions ("best X for Y", "X alternatives").
• Review directories and marketplaces — G2, Capterra, Product Hunt and their category equivalents. A complete, current listing is one of the cheapest citation sources available.
• Community discussions — Reddit threads and Stack Exchange answers are heavily cited for "what do people actually use" questions, precisely because they read as unsponsored.
• Video — YouTube results appear for how-to, setup and review queries, and are underweighted by most brands.
• Official product pages and documentation — strong for brand-specific and technical questions, weak for discovery, where a vendor claiming to be best is exactly what synthesis discounts.

Why do different engines give different answers to the same question?

Because each engine has its own index, its own query-rewriting behaviour, and its own synthesis style. Perplexity retrieves and cites aggressively and refreshes fast; Gemini grounds against Google's index, so its source mix correlates with Google rankings more than any other engine; ChatGPT's citations shift with its search partner data; Grok weights recent and X-native content unusually heavily.

The practical consequence is that the same question can name you in three answers out of five — and the two you lose are the actionable information. This is exactly why per-engine measurement matters more than a blended score: a large gap between your ChatGPT and Gemini numbers usually points at classic search visibility, while a Perplexity gap more often points at content freshness. AI Visibility Tracker records every result per engine for this reason; the engine-by-engine detail is in the track guides.

Does being cited require being linked?

No — and conflating the two is a common measurement error. An assistant can recommend a product by name with no link at all, and can cite a page about a product without naming the product in the answer text.

Both states are visibility, and they are worth distinguishing. A named-but-unlinked mention means the model knows you well enough to recommend you from its own synthesis; a linked-but-unnamed citation means a page about you was useful without you being the answer. Honest measurement checks the answer text and the structured citation list separately, then counts either as presence — which is what the Share of AI Answer formula does.

Can I influence which sources an engine chooses?

Indirectly and probabilistically. You cannot add a page to an engine's retrieval set, but you can make yourself present, current and quotable on the pages it already retrieves — and you can remove the obstacles that make your own pages unciteable, like crawler blocks, JS-only rendering, and product descriptions that never state plainly what the product is.

The gap analysis does the prioritising for you: the domains repeatedly cited on prompts where you are absent are a ranked list of places an engine already trusts for your exact question. Working that list is usually far cheaper than trying to outrank anyone. How to get cited by ChatGPT covers the concrete moves.